OpenAI’s coding agents now finish nearly half of the medium-length research jobs they’re handed, with no human stepping in. In a report published November 18, titled “Research acceleration: The jagged frontier,” the company says its agents completed 46% of research tasks estimated to take a human two to four hours. Back in March, that number was 23%. For longer jobs, four to eight hours of human work, the no-help success rate climbed from 13% to 35%.
Those numbers come from OpenAI’s own internal logs, and the company is the first to say they’re estimates rather than an outside audit. Still, they’re the clearest public window yet into how the lab that builds ChatGPT actually builds its models. Anthropic said much the same thing in a stock filing last week: its AI now speeds up its own research.
The headline hides a split.
What OpenAI coding agents can finish, and what they can’t
OpenAI’s core measure is zero-intervention success: an agent finished a delegated task without a person stepping in, as recorded in the company’s internal logs. The headline rates look strong. Underneath them, the picture breaks apart by task type.
Between August 1 and October 18, agents succeeded 79% of the time at monitoring training runs for problems, and 72% answering technical questions. They did well at communicating (69%) and analyzing (58%). Then the curve drops: launching a training run worked 35% of the time, setting up evaluation and reinforcement-learning environments hit 29%, and writing code for inference and serving stacks came in at just 25%.
“Agents now complete nearly half of two-to-four-hour research tasks on their own. But humans are still central to AI development,” the report says. OpenAI warns the weak categories could become bottlenecks to overall research gains, even as the headline numbers climb.
Researchers are running agents around the clock
Use is rising as fast as capability. OpenAI measures agent effort in “agent workdays,” eight hours of runtime each. That figure has jumped from 0.5 per researcher per day to 5.7, an elevenfold increase. On pretraining and posttraining teams, agent runtime averaged roughly eight workdays per researcher per calendar day.
Spending tells the same story. At API list prices, the median researcher’s daily agent use reached $2,400 in October, up 62% from $1,500 in September. At the 90th percentile it hit $23,000 a day, up 105% from $11,000. As of September, the API-equivalent value of agent use was about twice OpenAI’s non-equity pay for human researchers.
Agents don’t work one at a time anymore. By mid-October, more than two-thirds of the researchers studied averaged a daily peak of at least four agents running at once, and more than 10% averaged at least sixteen. The report names Codex and Dots as the internal tools driving the surge, according to The Decoder’s coverage.
The survey: researchers would trade colleagues for agents
Numbers only say so much about how much people value a tool. So OpenAI asked its own researchers a blunt hypothetical. From late August to early September, it surveyed 101 research staff: suppose your team lost access to all coding agents better than GPT-5.2 for the next three months. What would you trade to get them back?
The median answers were striking. Researchers said they’d give up 40% of team headcount, or 30% of team compute, to keep frontier coding agents for three months. Asked how much compute they’d trade if the question were posed in January 2027, the median answer rose to 50%.
A separate survey in late October put it more plainly. Asked whether they’d rather work with OpenAI’s latest internal model or a mid-career human research engineer helped by a 2024 model such as o1-preview, 27 of the 28 respondents chose or leaned toward the latest internal model. The big ideas behind GPT-6 Astra and GPT-6.2 Astra still came from humans, The Decoder reports, except for some pretraining insights found with Codex.
Why this matters
Two frontier labs have now said, in their own words, that AI is speeding up AI research. OpenAI’s version comes with receipts: success rates, spending, surveys, and a named weak spot. That’s what makes the report worth reading closely rather than skimming for the 46%.
First, the bottleneck matters more than the headline. A lab where agents write a quarter of serving code correctly still needs its best engineers on the systems that ship models to users. Progress in watching and talking won’t remove that constraint on its own.
Second, the money is now impossible to ignore. When a median researcher burns $2,400 a day in agent compute at list prices, and the top tenth burns $23,000, coding agents have stopped being a side tool. They’re a line item bigger than the salary of the person sitting at the desk, at least before equity.
None of this is a clean trend line. OpenAI says researchers changed which tasks they handed over as agents improved, so part of the climb may reflect better task-picking rather than better agents. And the company argues labs should be required to track and publicly report indicators of recursive self-improvement, the process by which AI speeds up the development of future AI systems, while stressing that publishing the numbers is not a plan to pursue it. The numbers are now on the record, and the next update will be judged against them.
Frequently asked questions
How often do OpenAI’s coding agents finish research tasks on their own?
For tasks estimated to take a human two to four hours, OpenAI’s coding agents now succeed with no human help 46% of the time, up from 23% in March 2026. For four-to-eight-hour tasks, the rate rose from 13% to 35%.
What are OpenAI coding agents best and worst at?
They’re best at watching work: 79% success monitoring training runs and 72% answering technical questions. They’re weakest at building: 25% writing inference and serving code, and 29% setting up evaluation and reinforcement-learning environments.
How much do OpenAI researchers spend on coding agents?
At API list prices, the median researcher used about $2,400 a day on agentic AI tools in October 2026, up 62% from September. The 90th-percentile researcher hit $23,000 a day.
What is recursive self-improvement?
It’s the process by which AI systems speed up the development of future AI systems. OpenAI says labs should track and publicly report indicators of it, while stressing that reporting on it is not a plan to pursue full recursive self-improvement.
Sources: OpenAI, “Research acceleration: The jagged frontier”; The Decoder, “OpenAI’s own researchers now say they value their AI tools more than nearly half their colleagues”; AI Weekly, “OpenAI Says Coding Agents Finish 46% of 2-4 Hour Research Tasks”; Superpower Daily, “OpenAI estimates coding agents finish 46% of its two-to-four-hour research tasks”; AI Daily Grind.
